This page is not an argument. It is a reading list, offered in the voice of a working research librarian: the long line of philosophers, mathematicians, physicists, physiologists, engineers, and neuroscientists who have tried, for roughly one thousand years, to say precisely what a mind does when it works out something it cannot directly see. Every claim below belongs to someone else, and is cited to them.
The word inference is older than the word science. Aristotle formalised deductive reasoning in the Prior Analytics (c. 350 BCE), giving the Western tradition its first grammar of what it means to derive a conclusion from premises. His syllogisms did not yet address uncertain conclusions from partial evidence — that would take much longer.
Ibn al-Haytham (Alhazen), working in Cairo around 1040 CE, wrote the seven-volume Kitāb al-Manāẓir — the Book of Optics. In it he argued that vision is not a passive receipt of light but a process of judgement: the mind actively supplies the missing information a retinal image cannot provide on its own. He is often credited with the first modern statement of the scientific method, and — as the physicist Jim Al-Khalili and the historian A. Mark Smith have both argued — with the first explicit account of perception as inference.
Francis Bacon's Novum Organum (1620) proposed that knowledge should be built up from particular observations to general laws. The trouble with this programme was posed most sharply, more than a century later, by David Hume.
Hume's challenge — that no finite set of observations can logically guarantee a general conclusion — is now called the problem of induction, and it remains open. It is the reason every serious theory of inference since Hume has been probabilistic rather than deductive: not what must be true, but how confident should I be. Immanuel Kant's Kritik der reinen Vernunft (1781) argued that the mind supplies the categories (space, time, causation) that make experience intelligible at all — a position later predictive-processing theorists would recognise as ancestral to their own.
Blaise Pascal and Pierre de Fermat, in a 1654 correspondence about a gambling puzzle, founded the mathematical theory of probability. A century later the Reverend Thomas Bayes wrote the paper that named an entire school of statistics.
Bayes's essay, published posthumously by Richard Price in 1763, gives the rule that would later be called Bayes's theorem — how to update the probability of a hypothesis in the light of new evidence. Pierre-Simon Laplace independently rediscovered and vastly extended the result; his Théorie analytique des probabilités (1812) is the true founding document of Bayesian inference as a working method for physics and astronomy. Andrey Kolmogorov's Grundbegriffe der Wahrscheinlichkeitsrechnung (1933) put probability on the modern measure-theoretic footing it still has today.
The idea that perception itself is a kind of inference — that seeing is not passive reception but an act of hypothesis-making — is not a modern invention. Hermann von Helmholtz named it in 1867.
Helmholtz's phrase — unbewusster Schluss, unconscious inference — is the direct ancestor of every twenty-first-century theory that treats the brain as a prediction machine. Every modern textbook of predictive processing opens with him.
Kenneth Craik, a Scottish psychologist who died at 31, wrote a small book in 1943 that is now regarded as the founding text of cognitive science.
Craik gave the twentieth century the concept of an internal model — a small, world-matching structure that the brain builds and updates. Every subsequent theory of the predictive brain owes him a citation.
The mid-twentieth century turned inference into engineering. Four papers, all published within seventeen years of each other, gave the field its modern grammar.
The physicist Richard Cox proved in 1946 that any consistent measure of plausibility — of "how much a rational agent should believe X given Y" — must obey the rules of probability. Probability, on this reading, is not a description of frequencies in the world; it is the logic of partial belief.
Edwin T. Jaynes spent his career building on this insight. His posthumous Probability Theory: The Logic of Science (2003) is the definitive statement of the position: that Bayesian inference is not one statistical method among many but the unique consistent way to reason from evidence to conclusion.
Two papers, seven years apart, brought Helmholtz's unconscious inference into the language of modern computational neuroscience.
The mathematician David Mumford, in 1992, proposed that the layered structure of the neocortex could be understood as a hierarchy in which higher levels send predictions down and lower levels send prediction errors up. Rajesh P. N. Rao and Dana H. Ballard turned this into a working model of visual cortex in 1999, giving predictive coding its now-canonical mathematical form and — importantly — an experimental prediction that was subsequently confirmed: that visual neurons should respond less strongly to expected stimuli than to unexpected ones.
By the mid-2000s the idea that the brain performs approximate Bayesian inference had moved from a philosopher's conjecture to a working research programme. David C. Knill and Alexandre Pouget's 2004 review in Trends in Neurosciences gave the position its name.
Karl Friston, at University College London, proposed in the mid-2000s a principle intended to unify perception, action, and learning under a single mathematical quantity: the variational free energy. The claim, briefly, is that any self-organising system that persists must act as if it is minimising a bound on the surprise of its own sensations — and that this single imperative can be shown, formally, to give rise to perception, action, learning, and attention as different facets of one computation.
The mature synthesis is the 2022 MIT Press textbook by Thomas Parr, Giovanni Pezzulo and Karl Friston. It is the reference every serious implementation — including UNI — cites first.
These authors, working in philosophy, cognitive science, and the sciences of mind, are useful companions for anyone reading the primary literature above. Each takes a different route into the same territory — the shared claim, in one form or another, is that living things persist by predicting and acting.
His Surfing Uncertainty (2016) is the standard-bearer for the philosophical reading of predictive processing. The Experience Machine (2023) is the general-audience follow-up.
The Predictive Mind (Oxford, 2013) is the rigorous, book-length case for the prediction-error framework as a theory of perception, agency, and self.
Being You (2021) reads predictive processing back into the question of consciousness itself — argued cautiously and empirically, not mystically.
Information Theory, Inference, and Learning Algorithms (Cambridge, 2003) is the friendliest bridge from Shannon and Bayes to modern probabilistic machine learning. The full book is legally free on the author's site.
Author of Probabilistic Reasoning in Intelligent Systems (1988) and Causality (2000/2009). His argument is that pure Bayesian conditioning is not enough — a full science of inference must also handle interventions, counterfactuals, and causal structure.
Reinforcement Learning: An Introduction (2nd ed., MIT Press, 2018). The canonical textbook for the learning-by-consequences tradition, which shares much machinery with active inference — and disagrees usefully on the rest.
Friston's principal co-authors on the 2022 MIT Press textbook. Pezzulo (ISTC-CNR Rome) has been the field's most consistent voice for a computational-cognitive reading of active inference; Parr has led its formal development.
Artificial Intelligence: A Modern Approach (with Peter Norvig, 4th ed., 2020) is the textbook that framed for two generations of engineers how inference — Bayesian and otherwise — is used inside working AI systems.
Ordered roughly by chronological entry point rather than by importance. Every item is a primary source or a well-regarded textbook, not a summary.
Nothing on this page belongs to us. UNI is not a discovery; it is an implementation of ideas that already have a nine-hundred-year paper trail. That is deliberate. If the mathematics is real, one small transparent piece of software should be able to obey it, step by step, on a laptop, and let anyone check the arithmetic.
What we contribute is care, reproducibility, and radical honesty about what is verified and what is not. Every claim on the labs pages of this site is tagged with an evidence class and a command a reader can run to check it. Every honesty banner names what the software is not. Everything on this page is somebody else's idea; we just do our best to implement it faithfully.
For the mathematical core, see our unrefereed preprint on the free-energy formulation:
Polzin et al. (2026). "An Organic Operator and AI Operator Collaborative Review of
Active Inference Free Energy Minimization."
doi:10.5281/zenodo.19785799
· Layer 2 expert review pending.
This bibliography is intentionally partial — a starting map, not a definitive genealogy. The field is broader than any one thread. If you find a citation here that is wrong, out of date, or that has a stronger primary source, we would like to know: the goal of the page is that it be right, not that it be ours. Corrections and additions are welcome via the site's GitHub, linked from About.
Last revised 2026-07-08. All citations refer to primary sources; where a DOI or an open link is available it is included.